Real-time simulation method and system for collision simulation analysis

Abstract

The application provides a real-time simulation method and system for collision simulation analysis. The real-time simulation method for collision simulation analysis comprises the following steps: S1, establishing a whole vehicle collision simulation analysis model according to the whole vehicle collision test working condition requirement; S2, defining the design variable and value range, test design sampling mode and sampling quantity based on the whole vehicle collision simulation analysis model, establishing a training sample simulation model and generating a corresponding training sample input data set; S3, completing the simulation analysis of the training sample simulation model and obtaining a simulation result; and S4, establishing a mapping relationship according to the simulation result, inputting a prediction sample input data set, and obtaining a prediction result to generate a prediction curve according to the established mapping relationship. The real-time simulation method and system for collision simulation analysis provided by the application save a large amount of software and hardware resources, greatly improve the simulation optimization efficiency, and significantly accelerate the research and development progress.

Description

Technical Field

[0001] This invention relates to the field of computer-aided design for automobiles, and specifically to a real-time simulation method and system for collision simulation analysis. Background Technology

[0002] A complete vehicle crash simulation analysis model often contains millions of mesh elements. Therefore, completing a single vehicle crash analysis requires significant hardware and software resources. Taking current mainstream new car development projects in China as an example, it typically requires a 128-core CPU, using a commercial crash solver, and continuous calculations for over 10 hours to obtain the vehicle crash simulation analysis results. Crash simulation calculations are extremely time-consuming, requiring not only hardware but also software resources. Furthermore, the development of a new car model often necessitates dozens of crash simulation scenarios. In addition, to obtain a body structure that meets crash performance requirements, simulation engineers generally need to perform dozens to hundreds of crash iteration calculations for each crash analysis scenario. For OEMs, this means investing substantial funds annually in leasing crash simulation software, building high-performance computer clusters, and hiring large engineering teams to meet the crash simulation needs of new car development.

[0003] Generally, after simulation engineers complete the construction and debugging of the basic simulation model for a specific crash analysis scenario, all subsequent iterative optimization work is based on this model. To ensure the simulation results meet performance targets or achieve optimal performance for the corresponding scenario, simulation engineers need to perform extensive manual iterations based on the basic simulation model. This process is time-consuming and labor-intensive, and may not necessarily yield the optimal solution that meets performance targets. Furthermore, the evaluation of crash simulation performance involves numerous aspects, including the simulation model's energy curve, the deformation process of key components of the vehicle structure, and the intrusion amount and velocity at key locations. Simulation engineers need to run the simulation result model of each calculation scheme using CAE post-processing tools to obtain the corresponding performance evaluation results. For simulation engineers, quickly obtaining the performance evaluation results of the whole-vehicle crash simulation analysis model can significantly improve development speed, thereby substantially reducing the development costs of new vehicle models. Summary of the Invention

[0004] The purpose of this invention is to provide a real-time simulation method and system for collision simulation analysis. The real-time simulation method provided by this invention predicts the performance curves and deformation profiles of collision simulation analysis conditions through dimensionality reduction algorithms and machine learning algorithms. Using a small amount of training sample data, a machine learning training model is established, after which all results within the design space can be simulated in real time. This invention is universally applicable to performance optimization of various collision simulation analysis conditions and also has strong guiding significance and reference value for performance optimization of other nonlinear simulation conditions.

[0005] To achieve the above objectives, this invention provides a real-time simulation method for collision simulation analysis, comprising:

[0006] Step S1: Establish a vehicle collision simulation analysis model according to the requirements of the vehicle collision test conditions;

[0007] Step S2: Based on the whole vehicle collision simulation analysis model established in Step S1, define the design variables and their value range, the experimental design sampling method and the number of samples, establish the training sample simulation model and generate the corresponding training sample input dataset.

[0008] Step S3: Complete the simulation analysis of the training sample simulation model established in step S2 and obtain the simulation results. The simulation results include: performance curve data of key positions or deformation contour data of key parts at the final moment.

[0009] Step S4: Establish a mapping relationship based on the simulation results obtained in step S3, input the prediction sample input dataset, and obtain the prediction results based on the established mapping relationship to generate a prediction curve. The generated prediction curve includes: the performance curve of the generated prediction sample or the final deformation profile of the generated prediction sample.

[0010] In one embodiment, the performance curve for generating the predicted sample includes:

[0011] Step S411: Extract performance curve data from key locations in the simulation results and generate a performance curve dataset;

[0012] Step S412: Perform data decomposition and dimensionality reduction on the performance curve dataset to obtain the first dimensionality-reduced feature matrix dataset;

[0013] Step S413: Construct the mapping relationship between the training sample input dataset and the first dimensionality-reduced feature matrix dataset;

[0014] Step S414: Input the prediction sample input dataset, and generate the second dimensionality-reduced feature matrix dataset corresponding to the prediction sample input dataset according to the mapping relationship constructed in step S413;

[0015] Step S415: Reconstruct the second dimensionality-reduced feature matrix dataset to complete data recovery and obtain the performance curve dataset of the predicted samples;

[0016] Step S416: Generate the performance curve of the prediction sample based on the obtained performance curve dataset of the prediction sample.

[0017] In one embodiment, step S412 includes:

[0018] Step S4121: Perform SVD singular value decomposition on the performance curve dataset Y1 to obtain feature matrices U1, S1, and V1. T The formula is as follows:

[0019] Y1=U1·S1·V1 T ;

[0020] Where Y1 is an n×p1 dimensional performance curve dataset, p1 is the number of output nodes of the performance curve at key positions, n is the number of training samples, U1 is an n×n unitary matrix, and V1 is a multiplicative matrix. T V1 is a p1×p1 unitary matrix, the conjugate transpose of V1, and S1 is a positive semi-definite n×p1 diagonal matrix. The elements s on the diagonal of S1 are... i These are the singular values ​​of Y1;

[0021] Step S4122: Perform dimensionality reduction on the performance curve dataset Y1 and generate the dimensionality-reduced performance curve dataset Y1. reduce The formula is as follows:

[0022] Y1 reduce =U1 reduce ·S1 reduce ·V1 T reduce ;

[0023] Among them, U1 reduce ·S1 reduce The dataset is the first dimensionality-reduced feature matrix dataset, with a dimensionality reduction dimension of L1 and L1 = n, V1 T reduce Principal component feature matrix, selected from V1 T The first L1 vectors.

[0024] In one embodiment, step S413 includes: constructing a training sample input dataset X and a first dimensionality-reduced feature matrix dataset U1 using a machine learning algorithm. reduce ·S1 reduce The mapping relationship f1 is given by the following formula:

[0025]

[0026] Where X is an n×m dimensional training sample input dataset, n is the number of training samples, and m is the number of design variables;

[0027] Step S414 includes: inputting the prediction sample input dataset XN, and generating the second dimensionality-reduced feature matrix dataset U1N corresponding to the prediction sample input dataset XN according to the mapping relationship f1. reduce ·S1N reduce The formula is as follows:

[0028]

[0029] Where XN is a k×m dimensional input dataset of prediction samples, k is the number of prediction samples, and m is the number of design variables;

[0030] Step S415 includes: processing the second dimensionality-reduced feature matrix dataset U1N reduce ·S1N reduce SVD reconstruction is performed to complete data recovery, obtaining the performance curve dataset Y1N of the k×p1 dimensional prediction samples, as shown in the following formula:

[0031] Y1N=U1N reduce ·S1N reduce ·V1 T reduce .

[0032] In one embodiment, generating the final deformed contour of the predicted sample in step S4 includes:

[0033] Step S421: Extract the final deformation contour data of key parts from the simulation results and generate a deformation contour dataset;

[0034] Step S422: Perform dimensionality reduction on the deformed contour dataset and generate the first deformed contour dimensionality-reduced dataset;

[0035] Step S423: Construct the mapping relationship between the training sample input dataset and the first deformed contour dimensionality reduction dataset;

[0036] Step S424: Input the prediction sample input dataset, and generate the second deformed contour dimensionality reduction dataset corresponding to the prediction sample input dataset according to the mapping relationship constructed in step S423;

[0037] Step S425: Reconstruct the second deformed contour dimensionality reduction dataset to obtain the deformed contour dataset of the predicted samples;

[0038] Step S426: Generate the final deformed contour of the predicted sample based on the obtained deformed contour dataset of the predicted sample.

[0039] In one embodiment, step S422 includes:

[0040] Step S4221: Generate covariance matrix data COV based on the deformed contour dataset Y2 Y2 The formula is as follows:

[0041] COV Y2 =Y2 T ·Y2;

[0042] Where Y2 is an n×p2 dimensional deformable contour dataset, n is the number of training samples, and p2 is the number of output nodes of the deformable contour;

[0043] Step S4222: Perform Singular Value Decomposition on the covariance matrix data using SVD. Y2 Perform data decomposition to obtain feature matrices U2, S2, and V2. T The formula is as follows:

[0044] COV Y2 =U2·S2·V2 T ;

[0045] Where U2 is a p2×p2 unitary matrix, V2 T S2 is a p2×p2 unitary matrix that is the conjugate transpose of V2, and S2 is a positive semi-definite p2×p2 diagonal matrix. The elements S on the diagonal of S2 are... i For COV Y2 The singular values;

[0046] Step S4223: Perform dimensionality reduction processing on the deformed contour dataset Y2 to obtain the first deformed contour dimensionality-reduced dataset Y2. reduce The formula is as follows:

[0047] Y2 reduce =Y2·U2 reduce ;

[0048] The dimension reduction dimension is L2, where L2 > n;

[0049] U2 reduce The principal component feature matrix is ​​selected from the first L2 vectors of U2.

[0050] In one embodiment, step S423 includes: using a machine learning algorithm to establish a training sample input dataset X and a first deformed contour dimensionality reduction dataset Y2. reduce The mapping relationship f2 is given by the following formula:

[0051]

[0052] Where X is an n×m dimensional training sample input dataset, n is the number of training samples, and m is the number of design variables;

[0053] Step S424 includes: inputting the prediction sample input dataset XN, and generating the second deformed contour dimensionality reduction dataset Y2N corresponding to the prediction sample input dataset XN according to the mapping relationship f2. reduce The formula is as follows:

[0054]

[0055] Where XN is a k×m dimensional input dataset of prediction samples, k is the number of prediction samples, and m is the number of design variables;

[0056] Step S425 includes: reconstructing the dimensionality reduction dataset Y2N of the second deformed contour using SVD. reduce To perform data recovery, obtain the deformed contour dataset Y2N of the k×p2 dimensional predicted samples, as shown in the following formula:

[0057] Y2N=U2N reduce ·S2N reduce ·V2 T reduce .

[0058] In one embodiment, the deformable contour dataset is: deformable contour datasets for the x-axis, y-axis, and z-axis generated respectively based on the three-dimensional coordinate data (x, y, z) of the final deformable contours of the key parts extracted from the simulation results, including: an n×p2 dimensional deformable contour dataset Y2x for the x-axis, an n×p2 dimensional deformable contour dataset Y2y for the y-axis, and an n×p2 dimensional deformable contour dataset Y2z for the z-axis.

[0059] Before executing step S426, substitute the deformed contour dataset Y2x of the x-axis, the deformed contour dataset Y2y of the y-axis, and the deformed contour dataset Y2z of the z-axis into Y2 respectively, and repeat steps S422 to S425 until the deformed contour dataset Y2Nx of the x-axis, the deformed contour dataset Y2Ny of the y-axis, and the deformed contour dataset Y2Nz of the z-axis of the predicted sample are obtained.

[0060] Step S426 includes: converting the deformation contour dataset Y2Nx of the x-axis of the predicted sample, the deformation contour dataset Y2Ny of the y-axis of the predicted sample, and the deformation contour dataset Y2Nz of the z-axis of the predicted sample into three-dimensional coordinate data of the predicted sample, so as to generate the final deformation contour of the predicted sample.

[0061] In one embodiment, the design variables include the thickness parameters and variation range of the B-pillar reinforcement plate and the B-pillar reinforcement plate cladding.

[0062] This invention provides a real-time simulation system for collision simulation analysis, including a controller and a memory. The memory stores a computer program, and the controller executes the computer program to implement the real-time simulation method for collision simulation analysis as described above.

[0063] The present invention has the following beneficial effects:

[0064] Within a defined design space, the real-time simulation method for collision simulation analysis provided by this invention establishes a mapping relationship between design variables and simulation results (such as performance curves and deformation profiles) through dimensionality reduction and machine learning algorithms. Engineers can then quickly obtain the analysis results for any sample point within the design space, achieving a real-time simulation effect. Since the analysis results predicted by real-time simulation do not require calling a solver for calculation, significant hardware and software resources are saved, and the development process is significantly accelerated. This method greatly improves simulation optimization efficiency, especially when the computational load of the simulation model is large, where the advantages of this algorithm are even more apparent. This invention is also applicable to all vehicle collision analysis scenarios, including but not limited to frontal collisions, frontal pole impacts, side collisions, side pole impacts, rear collisions, roof crush, and door intrusion. Attached Figure Description

[0065] Figure 1 This is a flowchart illustrating a real-time simulation method for collision simulation analysis according to an embodiment of the present invention.

[0066] Figure 2 This is a schematic diagram of the performance curve prediction process according to an embodiment of the present invention;

[0067] Figure 3 A schematic diagram of the deformation contour prediction process according to an embodiment of the present invention;

[0068] Figure 4 A schematic diagram of a simulated side-impact collision of a vehicle according to an embodiment of the present invention;

[0069] Figure 5 A schematic diagram of the design variables and intrusion measurement point locations according to an embodiment of the present invention;

[0070] Figure 6 Time-intrusion curves of H-point measurement points for sampling models with DV01 = 1.80 mm and DV02 = 1.80 mm;

[0071] Figure 7 A comparison of the deformation profiles of the B-pillar reinforcement plate at time zero and time final.

[0072] Figure 8 A comparison of the time-intrusion curves at the H-point measurement points for sampling models with DV01 = 2.05 mm and DV02 = 1.95 mm;

[0073] Figure 9 A comparison of the time-intrusion curves at the H-point measurement points for sampling models with DV01 = 2.10 mm and DV02 = 2.35 mm;

[0074] Figure 10A comparison of the time-intrusion curves at the H-point measurement points for sampling models with DV01 = 2.40 mm and DV02 = 1.80 mm;

[0075] Figure 11 A comparison of the time-intrusion curves at the H-point measurement points for sampling models with DV01 = 2.50 mm and DV02 = 2.50 mm;

[0076] Figure 12 A comparison of the final deformation profiles of the B-pillar reinforcement plate for sampling models with DV01 = 2.05 mm and DV02 = 1.95 mm;

[0077] Figure 13 A comparison of the final deformation profiles of the B-pillar reinforcement plate for sampling models with DV01 = 2.10 mm and DV02 = 2.35 mm;

[0078] Figure 14 A comparison of the final deformation profiles of the B-pillar reinforcement plate for sampling models with DV01 = 2.40 mm and DV02 = 1.80 mm;

[0079] Figure 15 Comparison of the final deformation profiles of the B-pillar reinforcement plate for sampling models with DV01 = 2.50 mm and DV02 = 2.50 mm. Detailed Implementation

[0080] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0081] like Figure 1 As shown, this invention provides a real-time simulation method for collision simulation analysis, including:

[0082] Step S1: Establish a vehicle collision simulation analysis model according to the requirements of the vehicle collision test conditions. It should be understood that the vehicle collision simulation analysis model can accurately reflect the collision performance of the vehicle model under the corresponding conditions. If there is corresponding vehicle collision test data for the corresponding collision conditions, it is necessary to ensure that the error between the result data calculated by the vehicle collision simulation analysis model and the vehicle collision test data is within the allowable range.

[0083] Step S2: Based on the vehicle collision simulation analysis model established in Step S1, define the design variables and their value ranges, the experimental design sampling method and the number of samples, establish a training sample simulation model, and generate the corresponding training sample input dataset. Preferably, in this embodiment, the design variables include the thickness parameters and variation range of the B-pillar reinforcement plate and the B-pillar reinforcement plate cladding (i.e., the B-pillar inner panel mentioned below), see [link to relevant documentation]. Figure 5 .

[0084] Step S3: Complete the simulation analysis of the training sample simulation model established in step S2 and obtain the simulation results. The simulation results include: performance curve data at key locations or final deformation profile data of key components.

[0085] Step S4: Establish a mapping relationship based on the simulation results obtained in Step S3, input the prediction sample input dataset, and obtain the prediction results according to the established mapping relationship to generate a prediction curve. The generated prediction curve includes: the performance curve of the generated prediction sample or the final deformation profile of the generated prediction sample.

[0086] Furthermore, in step S4 described above, generating the performance curve of the predicted sample includes:

[0087] Step S411: Extract performance curve data from key locations in the simulation results and generate a performance curve dataset;

[0088] Step S412: Perform data decomposition and dimensionality reduction on the performance curve dataset to obtain the first dimensionality-reduced feature matrix dataset;

[0089] Step S413: Construct the mapping relationship between the training sample input dataset and the first dimensionality-reduced feature matrix dataset;

[0090] Step S414: Input the prediction sample input dataset, and generate the second dimensionality-reduced feature matrix dataset corresponding to the prediction sample input dataset according to the mapping relationship constructed in step S413;

[0091] Step S415: Reconstruct the second dimensionality-reduced feature matrix dataset to complete data recovery and obtain the performance curve dataset of the predicted samples;

[0092] Step S416: Generate the performance curve of the prediction sample based on the obtained performance curve dataset of the prediction sample.

[0093] Furthermore, step S412 includes:

[0094] Step S4121: Perform Singular Value Decomposition (SVD) on the performance curve dataset Y1 to obtain the feature matrices U1, S1, and V1.T The formula is as follows:

[0095] Y1=U1·S1·V1 T ;

[0096] Where Y1 is an n×p1 dimensional performance curve dataset, p1 is the number of output nodes of the performance curve at key positions, n is the number of training samples, U1 is an n×n unitary matrix, and V1 is a multiplicative matrix. T V1 is a p1×p1 unitary matrix, the conjugate transpose of V1, and S1 is a positive semi-definite n×p1 diagonal matrix. The elements s on the diagonal of S1 are... i These are the singular values ​​of Y1.

[0097] Step S4122: Perform dimensionality reduction on the performance curve dataset Y1 and generate the dimensionality-reduced performance curve dataset Y1. reduce The formula is as follows:

[0098] Y1 reduce =U1 reduce ·S1 reduce ·V1 T reduce ;

[0099] Among them, U1 reduce ·S1 reduce V1 is the first dimensionality-reduced feature matrix dataset, with a dimensionality reduction dimension of L1 and L1 = n; T reduce Principal component feature matrix, selected from V1 T The first L1 vectors are key intermediate variables for data dimensionality reduction and reconstruction.

[0100] Step S413 includes:

[0101] Using machine learning algorithms, we construct the training sample input dataset X and the first dimensionality-reduced feature matrix dataset U1. reduce ·S1 reduce The mapping relationship f1 is given by the following formula:

[0102]

[0103] Where X is an n×m dimensional training sample input dataset, n is the number of training samples, and m is the number of design variables.

[0104] Step S414 includes:

[0105] Input the prediction sample input dataset XN, and generate the second dimensionality-reduced feature matrix dataset U1N corresponding to the prediction sample input dataset XN according to the mapping relationship f1. reduce ·S1N reduce The formula is as follows:

[0106]

[0107] Where XN is a k×m dimensional input dataset of prediction samples, k is the number of prediction samples, and m is the number of design variables;

[0108] Step S415 includes:

[0109] For the second dimensionality-reduced feature matrix dataset U1N reduce ·S1N reduce SVD reconstruction is performed to complete data recovery, obtaining the performance curve dataset Y1N of the k×p1 dimensional prediction samples, as shown in the following formula:

[0110] Y1N=U1N reduce ·S1N reduce ·V1 T reduce .

[0111] Furthermore, the performance curve dataset Y1N of the predicted samples is transformed into performance curves for each predicted sample. In this embodiment, the performance curves are time-performance curves. Preferably, the performance curve dataset Y1N of the predicted samples is saved as a text file or a data file in Excel / CSV format, making it convenient for users to import it into CAE post-processing tools or other customized software to view curve changes or perform data comparisons.

[0112] Furthermore, in step S4 described above, generating the final deformed contour of the predicted sample includes:

[0113] Step S421: Extract the final deformation contour data of key parts from the simulation results and generate a deformation contour dataset;

[0114] Step S422: Perform dimensionality reduction on the deformed contour dataset and generate the first deformed contour dimensionality-reduced dataset;

[0115] Step S423: Construct the mapping relationship between the training sample input dataset and the first deformed contour dimensionality reduction dataset;

[0116] Step S424: Input the prediction sample input dataset, and generate the second deformed contour dimensionality reduction dataset corresponding to the prediction sample input dataset according to the mapping relationship constructed in step S423;

[0117] Step S425: Reconstruct the second deformed contour dimensionality reduction dataset to obtain the deformed contour dataset of the predicted samples;

[0118] Step S426: Generate the final deformed contour of the predicted sample based on the obtained deformed contour dataset of the predicted sample.

[0119] Furthermore, step S422 includes:

[0120] Step S4221: Generate covariance matrix data COV based on the deformed contour dataset Y2 Y2 The formula is as follows:

[0121] COV Y2 =Y2 T ·Y2;

[0122] Where Y2 is an n×p2 dimensional deformable contour dataset, n is the number of training samples, and p2 is the number of output nodes of the deformable contour.

[0123] Step S4222: Perform Singular Value Decomposition on the covariance matrix data using SVD. Y2 Perform data decomposition to obtain feature matrices U2, S2, and V2. T The formula is as follows:

[0124] COV Y2 =U2·S2·V2 T ;

[0125] Where U2 is a p2×p2 unitary matrix, V2 T S2 is a p2×p2 unitary matrix that is the conjugate transpose of V2, and S2 is a positive semi-definite p2×p2 diagonal matrix. The elements S on the diagonal of S2 are... i For COV Y2 The singular values ​​of .

[0126] Step S4223: Perform dimensionality reduction processing on the deformed contour dataset Y2 to obtain the first deformed contour dimensionality-reduced dataset Y2. reduce The formula is as follows:

[0127] Y2 reduce =Y2·U2 reduce ;

[0128] The dimension reduction dimension is L2, where L2 > n;

[0129] U2 reduce Principal component feature matrix, selected from the first L2 vectors of U2, is a key intermediate variable for data dimensionality reduction and reconstruction.

[0130] Step S423 includes:

[0131] The training sample input dataset X and the first deformed contour dimensionality reduction dataset Y2 are established using machine learning algorithms. reduce The mapping relationship f2 is given by the following formula:

[0132]

[0133] Where X is an n×m dimensional training sample input dataset, n is the number of training samples, and m is the number of design variables.

[0134] Step S424 includes:

[0135] Given the input dataset XN containing the prediction samples, and based on the mapping relationship f2, generate the second deformed contour dimensionality reduction dataset Y2N corresponding to the input dataset XN. reduce The formula is as follows:

[0136]

[0137] Where XN is a k×m dimensional input dataset of prediction samples, k is the number of prediction samples, and m is the number of design variables.

[0138] Step S425 includes:

[0139] The second deformed contour dimensionality reduction dataset Y2N was reconstructed using SVD. reduce To perform data recovery, obtain the deformed contour dataset Y2N of the k×p2 dimensional predicted samples, as shown in the following formula:

[0140] Y2N=U2N reduce ·S2N reduce ·V2 T reduce .

[0141] In this embodiment, the deformation contour dataset is: the deformation contour datasets of the x-axis, y-axis and z-axis generated respectively based on the three-dimensional coordinate data (x, y, z) of the final deformation contour of the key parts extracted from the simulation results, including: the n×p2-dimensional x-axis deformation contour dataset Y2x, the n×p2-dimensional y-axis deformation contour dataset Y2y, and the n×p2-dimensional z-axis deformation contour dataset Y2z.

[0142] Before executing step S426, substitute the deformed contour dataset Y2x of the x-axis, the deformed contour dataset Y2y of the y-axis, and the deformed contour dataset Y2z of the z-axis into Y2 respectively, and repeat steps S422 to S425 until the deformed contour dataset Y2Nx of the x-axis, the deformed contour dataset Y2Ny of the y-axis, and the deformed contour dataset Y2Nz of the z-axis of the predicted sample are obtained.

[0143] Step S426 includes: converting the deformation contour dataset Y2Nx (x-axis), Y2Ny (y-axis), and Y2Nz (z-axis) of the predicted sample into three-dimensional coordinate data of the predicted sample to generate the final deformation contour of the predicted sample. Preferably, in this embodiment, the three-dimensional coordinate data of the predicted sample is further converted into a model text file or post-processing format file corresponding to the simulation solver, making it convenient for users to import into CAE post-processing tools or other customized software to view the deformation contour.

[0144] To make the above-mentioned objectives, features and advantages of the present invention more apparent and understandable, a specific embodiment of the present invention will be described in further detail below. This embodiment is a performance optimization of a vehicle side collision analysis under working conditions.

[0145] In this embodiment, the real-time simulation method for collision simulation analysis includes:

[0146] Step S1: Establish a vehicle collision simulation analysis model according to the requirements of the vehicle collision test conditions. In this embodiment, following the company's internal collision modeling standard process, a simulation model of a certain vehicle's side collision condition is established in the CAE preprocessing software. For example... Figure 4 As shown, a moving deformable barrier impacts the test vehicle from the side perpendicularly at a speed of 50 km / h. The calculation time is 150 ms, with a time step of 2 ms. The established vehicle collision simulation analysis model will be solved using a collision simulation solver, and the calculation result file will be obtained. The collision result data, including vehicle body structure deformation and collision performance curves, will be obtained through CAE post-processing software.

[0147] Step S2: Based on the vehicle collision simulation analysis model established in Step S1, define the design variables and their value ranges, the experimental design sampling method and the number of samples, establish a training sample simulation model, and generate the corresponding n×m dimensional training sample input dataset X. In this embodiment, the initial values ​​and variation ranges of the design variables are shown in Table 1 below.

[0148] Table 1 Initial values ​​and range of variation of design variables

[0149]

[0150] A model with 9 training samples was established using the 2-factor, 3-level full factorial method. The number of training samples n = 9, the number of design variables m = 2, and the corresponding input dataset X for the training samples is as follows:

[0151]

[0152] Step S3: Complete the simulation analysis of the nine training sample simulation models established in Step S2 and obtain the simulation results. The simulation results include: performance curve data at key locations or final deformation profile data of key components.

[0153] Step S4: Establish a mapping relationship based on the simulation results obtained in Step S3, input the prediction sample input dataset, and obtain the prediction results according to the established mapping relationship to generate a prediction curve. The generated prediction curve includes: the performance curve of the generated prediction sample or the final deformation profile of the generated prediction sample.

[0154] In this embodiment, the performance curve examined is the time-intrusion curve of the H-point measurement point (node ​​number: 78250005) below the inner panel of the B-pillar. The location of the intrusion is shown in [reference needed]. Figure 5 The time-intrusion curves of the H-point measurement points for the sampling models with DV01 = 1.80 mm and DV02 = 1.80 mm are shown below. Figure 6 As shown. This time-intrusion curve includes 76 time points, so the number of output nodes p1 = 76.

[0155] The deformation profile examined is the final deformation profile of the B-pillar reinforcement plate. Specifically, the deformation profiles of the B-pillar reinforcement plate at time zero and time final for the sampling model with DV01 = 1.80 mm and DV02 = 1.80 mm are shown below. Figure 7 As shown in (a) and (b), the finite element model of the B-pillar reinforcement plate includes 15,866 mesh element nodes, so the number of nodes output for the deformation profile is p2 = 15,866.

[0156] Steps S411 to S416 are used to generate the time-invasiveness curve for predicting H-point measurements. For the specific prediction process, please refer to [link / reference needed]. Figure 2 .

[0157] Step S411: Extract performance curve data from key locations in the simulation results and generate a performance curve dataset. In this embodiment, the time-intrusion amount of the H-point measurement points in the simulation results is extracted, and a 9×76-dimensional performance curve dataset Y1 is generated, as follows:

[0158]

[0159] Step S412: Perform data decomposition and dimensionality reduction on the performance curve dataset to obtain the first dimensionality-reduced feature matrix dataset, specifically:

[0160] Step S4121: Perform SVD singular value decomposition on the performance curve dataset Y1 to obtain feature matrices U1, S1, and V1. T The formula is as follows:

[0161] Y1=U1·S1·V1 T ;

[0162] Where U1 is an n×n unitary matrix, V1 T V1 is a p1×p1 unitary matrix, the conjugate transpose of V1, and S1 is a positive semi-definite n×p1 diagonal matrix. The elements s on the diagonal of S1 are... i These are the singular values ​​of Y1. In this embodiment, n = 9, p1 = 76.

[0163] Step S4122: Perform dimensionality reduction on the performance curve dataset Y1 and generate the dimensionality-reduced performance curve dataset Y1. reduce The formula is as follows:

[0164] Y1 reduce =U1 reduce ·S1 reduce ·V1 T reduce ;

[0165] Among them, U1 reduce ·S1 reduce The dataset is the first dimensionality-reduced feature matrix dataset, with a dimensionality reduction dimension of L1 and L1 = n, V1 T reduce Principal component feature matrix, selected from V1 T The first L1 vectors. And L1 is much smaller than p1. In this embodiment, preferably, L1 = 9, then a 9×9 dimensional first reduced feature matrix dataset U1 is obtained. reduce ·S1 reduce as follows:

[0166]

[0167] Step S413: Construct the training sample input dataset X and the first dimensionality-reduced feature matrix dataset U1 using machine learning algorithms. reduce ·S1 reduce The mapping relationship f1 is given by the following formula:

[0168]

[0169] In this embodiment, machine learning training is performed by calling the Python machine learning toolkit scikit-learn (an open-source machine learning toolkit), using the Random Forest algorithm. Testing showed that the machine learning model corresponding to each output parameter meets the accuracy requirements.

[0170] Step S414: Input the prediction sample input dataset XN, and generate the second dimensionality-reduced feature matrix dataset U1N corresponding to the prediction sample input dataset XN according to the mapping relationship f1. reduce ·S1N reduce The formula is as follows:

[0171]

[0172] Where XN is a k×m dimensional input dataset of prediction samples, k is the number of prediction samples, and m is the number of design variables.

[0173] In this embodiment, four prediction samples within the design space were randomly selected for H-point measurement point time-intrusion curve prediction, so k=4 and m=2. The corresponding prediction sample input dataset XN is as follows:

[0174]

[0175] By using the mapping relationship f1, we obtain the second dimensionality-reduced feature matrix dataset U1N corresponding to XN, which is 4×9 dimensions. reduce ·S1N reduce as follows:

[0176]

[0177] Step S415: Process the second dimensionality-reduced feature matrix dataset U1N reduce ·S1N reduce SVD reconstruction is performed to complete data recovery, obtaining the performance curve dataset Y1N of the k×p1 dimensional prediction samples, as shown in the following formula:

[0178] Y1N=U1N reduce ·S1N reduce ·V1 T reduce .

[0179] In this embodiment, the performance curve dataset Y1N of the 4×76-dimensional prediction samples is obtained as follows:

[0180]

[0181] Step S416: Convert the performance curve dataset Y1N of the predicted samples into the invasion amount time-performance curve of each predicted sample, and save it as a text format or data format file such as Excel / CSV, so that users can import it into CAE post-processing tools or other customized software to view the curve changes or make data comparisons.

[0182] To verify the accuracy of the prediction results, this embodiment simultaneously calculated the simulation results of four prediction sampling models corresponding to the prediction sample input dataset XN, and extracted the corresponding H-point measurement point time-intrusion curves. A comparison of the time-intrusion curves extracted from the simulation models and those generated by the method of this invention is shown in the figure below. Figure 8-11 As shown. By Figure 8-11 As can be seen, the predicted results and the simulation results are very close.

[0183] Steps S421 to S426 are used to predict the final deformation profile of the B-pillar reinforcement plate at T = 150 ms. For the specific prediction process, please refer to [link to relevant documentation]. Figure 3 .

[0184] Step S421: Extract the deformation contours of key components from the simulation results of the training sample model. Based on the three-dimensional coordinate data (x, y, z) of the final deformation contours of the key components extracted from the simulation results, generate deformation contour datasets for the x-axis, y-axis, and z-axis respectively, including: an n×p² dimensional x-axis deformation contour dataset Y2x, an n×p² dimensional y-axis deformation contour dataset Y2y, and an n×p² dimensional z-axis deformation contour dataset Y2z. Since the B-pillar reinforcement plate includes 15866 mesh element nodes, the number of deformation contour output nodes p2 = 15866. Therefore, the x-axis deformation contour data of all sample points constitute a 9×15866 dimensional deformation contour dataset Y2x, the y-axis deformation contour data constitute a 9×15866 dimensional deformation contour dataset Y2y, and the z-axis deformation contour data constitute a 9×15866 dimensional deformation contour dataset Y2z.

[0185] The following describes steps S421 to S425 in detail, using the deformed contour dataset Y2x along the x-axis as an example. In this embodiment, the 9×15866-dimensional deformed contour dataset Y2x along the x-axis is as follows:

[0186]

[0187] Step S422: Perform dimensionality reduction on the deformed contour dataset Y2x along the x-axis to generate the first deformed contour dimensionality-reduced dataset. Step S422 includes:

[0188] Step S4221: Generate covariance matrix data COV based on the deformed contour dataset Y2x. Y2x The formula is as follows:

[0189] COV Y2x =Y2x T ·Y2x;

[0190] Step S4222: Perform Singular Value Decomposition on the covariance matrix data using SVD. Y2x Perform data decomposition to obtain feature matrices U2x, S2x, and V2x. T The formula is as follows:

[0191] COV Y2x =U²x·S²x·V²x T ;

[0192] Where U2x is a p2×p2 unitary matrix, V2x T S2x is a p2×p2 unitary matrix that is the conjugate transpose of V2x, and S2x is a positive semi-definite p2×p2 diagonal matrix. The elements S on the diagonal of S2x are... i For COV Y2x The singular value of p2. In this embodiment, p2 = 15866.

[0193] Step S4223: Perform dimensionality reduction processing on the deformed contour dataset Y2x along the x-axis to obtain the first deformed contour dimensionality-reduced dataset Y2x. reduce The formula is as follows:

[0194] Y2x reduce =Y2x·U2x reduce ;

[0195] Where the dimensionality reduction dimension is L2, L2>n; U2x reduce Principal component feature matrix, selected from the first L2 vectors of U2x, is a key intermediate variable for data dimensionality reduction and reconstruction.

[0196] In this embodiment, setting L2 = 100 yields a 9×100 dimensionality reduction dataset Y2x of deformed contours along the x-axis. reduce as follows:

[0197]

[0198] Step S423: Use machine learning algorithms to establish the first deformed contour dimensionality reduction dataset Y2x for the training sample input dataset X and x-axis. reduce The mapping relationship f2 is given by the following formula:

[0199]

[0200] In this embodiment, machine learning training is performed by calling the Python machine learning toolkit scikit-learn (an open-source machine learning toolkit), using the Random Forest algorithm. Testing showed that the machine learning model corresponding to each output parameter meets the accuracy requirements.

[0201] Step S424: Input the prediction sample input dataset XN, and generate the second deformed contour dimensionality reduction dataset Y2Nx corresponding to the prediction sample input dataset XN according to the mapping relationship f2. reduce The formula is as follows:

[0202]

[0203] Where XN is a k×m dimensional input dataset of prediction samples, k is the number of prediction samples, and m is the number of design variables.

[0204] In this embodiment, four prediction samples within the design space were randomly selected for predicting the deformed profile of the reinforced B-pillar, so k = 4 and m = 2. The corresponding prediction sample input dataset XN is as follows:

[0205]

[0206] By using the mapping relationship f2, we obtain the 4×100 dimensionality-reduced dataset Y2Nx corresponding to XN. reduce as follows:

[0207]

[0208] Step S425 includes: reconstructing the dimensionality reduction dataset Y2Nx of the second deformed contour using SVD. reduce To perform data recovery, obtain the deformed contour dataset Y2Nx of the k×p2 dimensional predicted samples, as shown in the following formula:

[0209] Y2Nx=U2Nx reduce ·S2Nx reduce V2x T reduce .

[0210] In this embodiment, the deformed contour dataset Y2Nx of the x-axis of the 4×15866-dimensional predicted samples is obtained as follows:

[0211]

[0212] Before executing step S426, it is necessary to substitute the deformed contour dataset Y2y of the y-axis and the deformed contour dataset Y2z of the z-axis into Y2 respectively, and repeat steps S422 to S425 until the deformed contour dataset Y2Nx of the x-axis, the deformed contour dataset Y2Ny of the y-axis, and the deformed contour dataset Y2Nz of the z-axis of the predicted sample are obtained.

[0213] Step S426 includes: converting the deformation contour dataset Y2Nx (x-axis), Y2Ny (y-axis), and Y2Nz (z-axis) of the predicted sample into three-dimensional coordinate data of the predicted sample to generate the final deformation contour of the predicted sample. Then, the three-dimensional coordinate data of the predicted sample is converted into a model text file or post-processing format file corresponding to the simulation solver, making it convenient for users to import into CAE post-processing tools or other customized software to view the deformation contour. To verify the accuracy of the prediction results, this case simultaneously calculated the simulation results of the four predicted sampling models corresponding to the input dataset XN of the predicted sample and extracted the deformation contour of the corresponding B-pillar reinforcement plate at the final moment. A comparison diagram of the final moment deformation contour of the B-pillar reinforcement plate extracted from the simulation model and the final moment deformation contour of the B-pillar reinforcement plate generated by the method of this invention is shown below. Figure 12-15 As shown. By Figure 12-15 As can be seen, the predicted results and the simulation results are very close.

[0214] This invention also provides a real-time simulation system for collision simulation analysis, including a controller and a memory. The memory stores a computer program, and the controller executes the computer program to implement the real-time simulation method for collision simulation analysis as described above.

[0215] The present invention has the following beneficial effects:

[0216] Within a defined design space, the real-time simulation method for collision simulation analysis provided by this invention establishes a mapping relationship between design variables and simulation results (such as performance curves and deformation profiles) through dimensionality reduction and machine learning algorithms. Engineers can then quickly obtain the analysis results for any sample point within the design space, achieving a real-time simulation effect. Since the analysis results predicted by real-time simulation do not require calling a solver for calculation, significant hardware and software resources are saved, and the development process is significantly accelerated. This method greatly improves simulation optimization efficiency, especially when the computational load of the simulation model is large, where the advantages of this algorithm are even more apparent. This invention is also applicable to all vehicle collision analysis scenarios, including but not limited to frontal collisions, frontal pole impacts, side collisions, side pole impacts, rear collisions, roof crush, and door intrusion.

[0217] The embodiments described in this specification adopt a progressive approach, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. The combination of technical features in this invention is not limited to the combinations described in this application; all technical features described in this invention can be freely combined or combined in any way, unless they contradict each other.

[0218] The embodiments described above are merely further illustrations of the present invention and are not intended to limit the present invention in any other way. The present invention may have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding modifications and changes based on the present invention, but all such modifications and changes should fall within the protection scope of the present invention.

Claims

1. A real-time simulation method for collision simulation analysis, characterized in that, include: Step S1: Establish a vehicle collision simulation analysis model according to the requirements of the vehicle collision test conditions; Step S2: Based on the whole vehicle collision simulation analysis model established in Step S1, define the design variables and their value range, the experimental design sampling method and the number of samples, establish the training sample simulation model and generate the corresponding training sample input dataset. Step S3: Complete the simulation analysis of the training sample simulation model established in step S2 and obtain the simulation results. The simulation results include: performance curve data of key positions or deformation contour data of key parts at the final moment. Step S4: Establish a mapping relationship based on the simulation results obtained in step S3, input the prediction sample input dataset, and obtain the prediction results based on the established mapping relationship to generate a prediction curve. The generated prediction curve includes: the performance curve of the generated prediction sample or the final deformation profile of the generated prediction sample. The step S4 of generating the final deformed contour of the predicted sample includes: Step S421: Extract the final deformation contour data of key parts from the simulation results and generate a deformation contour dataset; Step S422: Perform dimensionality reduction on the deformed contour dataset and generate the first deformed contour dimensionality-reduced dataset; Step S423: Construct the mapping relationship between the training sample input dataset and the first deformed contour dimensionality reduction dataset; Step S424: Input the prediction sample input dataset, and generate the second deformed contour dimensionality reduction dataset corresponding to the prediction sample input dataset according to the mapping relationship constructed in step S423; Step S425: Reconstruct the second deformed contour dimensionality reduction dataset to obtain the deformed contour dataset of the predicted samples; Step S426: Generate the final deformed contour of the predicted sample based on the obtained deformed contour dataset of the predicted sample. Step S422 includes: Step S4221: Generate covariance matrix data COV based on the deformed contour dataset Y2 Y2 The formula is as follows: COV Y2 = Y2 T ·Y2; Where Y2 is an n×p2 dimensional deformable contour dataset, n is the number of training samples, and p2 is the number of output nodes of the deformable contour; Step S4222: Perform Singular Value Decomposition on the covariance matrix data using SVD. Y2 Perform data decomposition to obtain feature matrices U2, S2, and V2. T The formula is as follows: COV Y2 = U2·S2·V2 T ; Where U2 is a p2×p2 unitary matrix, V2 T S2 is a p2×p2 unitary matrix that is the conjugate transpose of V2, and S2 is a positive semi-definite p2×p2 diagonal matrix. The elements on the diagonal of S2 are... For COV Y2 The singular values; Step S4223: Perform dimensionality reduction processing on the deformed contour dataset Y2 to obtain the first deformed contour dimensionality-reduced dataset Y2. reduce The formula is as follows: Y2 reduce = Y2·U2 reduce ; The dimension reduction dimension is L2, where L2 > n; U2 reduce The principal component feature matrix is ​​selected from the first L2 vectors of U2.

2. The real-time simulation method for collision simulation analysis according to claim 1, characterized in that, The performance curve for generating the predicted samples includes: Step S411: Extract performance curve data from key locations in the simulation results and generate a performance curve dataset; Step S412: Perform data decomposition and dimensionality reduction on the performance curve dataset to obtain the first dimensionality-reduced feature matrix dataset; Step S413: Construct the mapping relationship between the training sample input dataset and the first dimensionality-reduced feature matrix dataset; Step S414: Input the prediction sample input dataset, and generate the second dimensionality-reduced feature matrix dataset corresponding to the prediction sample input dataset according to the mapping relationship constructed in step S413; Step S415: Reconstruct the second dimensionality-reduced feature matrix dataset to complete data recovery and obtain the performance curve dataset of the predicted samples; Step S416: Generate the performance curve of the prediction sample based on the obtained performance curve dataset of the prediction sample.

3. The real-time simulation method for collision simulation analysis according to claim 2, characterized in that, Step S412 includes: Step S4121: Perform SVD singular value decomposition on the performance curve dataset Y1 to obtain feature matrices U1, S1, and V1. T The formula is as follows: Y1 = U1·S1·V1 T ; Where Y1 is an n×p1 dimensional performance curve dataset, p1 is the number of output nodes of the performance curve at key positions, n is the number of training samples, U1 is an n×n unitary matrix, and V1 is a multiplicative matrix. T V1 is a p1×p1 unitary matrix, the conjugate transpose of V1, and S1 is a positive semi-definite n×p1 diagonal matrix. The elements on the diagonal of S1... These are the singular values ​​of Y1; Step S4122: Perform dimensionality reduction on the performance curve dataset Y1 and generate the dimensionality-reduced performance curve dataset Y1. reduce The formula is as follows: Y1 reduce = U1 reduce ·S1 reduce ·V1 T reduce ; Among them, U1 reduce ·S1 reduce The dataset is the first dimensionality-reduced feature matrix dataset, with a dimensionality reduction dimension of L1 and L1=n, V1 T reduce Principal component feature matrix, selected from V1 T The first L1 vectors.

4. The real-time simulation method for collision simulation analysis according to claim 3, characterized in that, Step S413 includes: constructing a training sample input dataset X and a first dimensionality-reduced feature matrix dataset U1 using a machine learning algorithm. reduce ·S1 reduce mapping relationship f1 The formula is as follows: X U1 reduce ·S1 reduce ; Where X is an n×m dimensional training sample input dataset, n is the number of training samples, and m is the number of design variables; Step S414 includes: inputting the prediction sample input dataset XN, and according to the mapping relationship... f1 Generate the second dimensionality-reduced feature matrix dataset U1N corresponding to the input dataset XN of the prediction sample dataset. reduce ·S1N reduce The formula is as follows: XN U1N reduce ·S1N reduce ; Where XN is a k×m dimensional input dataset of prediction samples, k is the number of prediction samples, and m is the number of design variables; Step S415 includes: processing the second dimensionality-reduced feature matrix dataset U1N reduce ·S1N reduce Perform SVD reconstruction to complete data recovery and obtain... k×p1 The performance curve of the predicted samples for the dataset Y1N is given by the following formula: Y1N = U1N reduce ·S1N reduce ·V1 T reduce 。 5. The real-time simulation method for collision simulation analysis according to claim 1, characterized in that, Step S423 includes: using machine learning algorithms to establish a training sample input dataset X and a first deformed contour dimensionality reduction dataset Y2. reduce mapping relationship f2 The formula is as follows: X Y2 reduce ; Where X is an n×m dimensional training sample input dataset, n is the number of training samples, and m is the number of design variables; Step S424 includes: inputting the prediction sample input dataset XN, and according to the mapping relationship... f2 Generate the second deformed contour dimensionality reduction dataset Y2N corresponding to the input dataset XN of the prediction sample. reduce The formula is as follows: XN Y2N reduce ; Where XN is a k×m dimensional input dataset of prediction samples, k is the number of prediction samples, and m is the number of design variables; Step S425 includes: reconstructing the dimensionality reduction dataset Y2N of the second deformed contour using SVD. reduce To perform data recovery, obtain the deformed contour dataset Y2N of the k×p2 dimensional predicted samples, as shown in the following formula: Y2N = U2N reduce ·S2N reduce ·V2 T reduce 。 6. The real-time simulation method for collision simulation analysis according to claim 5, characterized in that, The deformable contour dataset is: based on the three-dimensional coordinate data (x, y, z) of the final deformable contour of the key parts extracted from the simulation results, the deformable contour datasets of the x-axis, y-axis and z-axis are generated respectively, including: the n×p2 dimension x-axis deformable contour dataset Y2x, the n×p2 dimension y-axis deformable contour dataset Y2y, and the n×p2 dimension z-axis deformable contour dataset Y2z. Before executing step S426, substitute the deformed contour dataset Y2x of the x-axis, the deformed contour dataset Y2y of the y-axis, and the deformed contour dataset Y2z of the z-axis into Y2 respectively, and repeat steps S422 to S425 until the deformed contour dataset Y2Nx of the x-axis, the deformed contour dataset Y2Ny of the y-axis, and the deformed contour dataset Y2Nz of the z-axis of the predicted sample are obtained. Step S426 includes: converting the deformation contour dataset Y2Nx of the x-axis of the predicted sample, the deformation contour dataset Y2Ny of the y-axis of the predicted sample, and the deformation contour dataset Y2Nz of the z-axis of the predicted sample into three-dimensional coordinate data of the predicted sample, so as to generate the final deformation contour of the predicted sample.

7. The real-time simulation method for collision simulation analysis according to claim 1, characterized in that, The design variables include the thickness parameters and variation range of the B-pillar reinforcement plate and the B-pillar reinforcement plate cladding.

8. A real-time simulation system for collision simulation analysis, characterized in that, It includes a controller and a memory, wherein the memory stores a computer program, and the controller executes the computer program to implement the real-time simulation method for collision simulation analysis as described in any one of claims 1-7.

Images